A fully multiple-criteria implementation of the Sobol' method for parameter sensitivity analysis

We present a novel rank-based fully multiple-criteria implementation of the Sobol variance-based sensitivity analysis approach that implements an objective strategy to evaluate parameter sensitivity when model evaluation involves several metrics of performance. The method is superior to single-criterion approaches while avoiding the subjectivity observed in "pseudo" multiple-criteria methods. Further, it contributes to our understanding of structural characteristics of a model and simplifies parameter estimation by identifying insensitive parameters that can be fixed to default values during model calibration studies. We illustrate the approach by applying it to the problem of identifying the most influential parameters in the Simple Biosphere 3 (SiB3) model using a network of flux towers in Brazil. We find 27-31 (out of 42) parameters to be influential, most (∼78%) of which are primarily associated with physiology, soil, and carbon properties, and that uncertainties in the physiological properties of the model contribute most to total model uncertainty in regard to energy and carbon fluxes. We also find that the second most important model component contributing to the total output uncertainty varies according to the flux analyzed; whereas morphological properties play an important role in sensible heat flux, soil properties are important for latent heat flux, and carbon properties (mainly associated with the soil respiration submodel) are important for carbon flux (as expected). These distinct sensitivities emphasize the need to account for the multioutput nature of land surface models during sensitivity analysis and parameter estimation. Applied to other similar models, our approach can help to establish which soil-plant-atmosphere processes matter most in land surface models of Amazonia and thereby aid in the design of field campaigns to characterize and measure the associated parameters. The approach can also be used with other sensitivity analysis procedures that compute at least two model performance metrics.